The Algorithm That Couldn’t Save Itself: What AI’s Biggest Drug Discovery Bets Reveal About the Future of Medicine

Image AI generated

Picture a boardroom in Oxford, 2020. Executives from Exscientia and Sumitomo Pharma are celebrating what the press will call a historic moment: the world’s first AI-designed drug to enter human clinical trials.

DSP-1181, a molecule for obsessive-compulsive disorder, had been conjured not by years of painstaking lab work but by an algorithm — in a fraction of the usual time.

The champagne was real. The hope was real.

Two years later, the drug was quietly discontinued.

This is the story of AI in pharma — not the version told in investor decks, but the messier, more instructive one — where breathtaking speed meets the stubborn complexity of human biology, and where the most important question isn’t can AI discover drugs faster? but does faster even matter if the drugs still don’t work?

A Revolution That Actually Happened (Sort Of)

The numbers are genuinely staggering. The AI drug discovery market is projected to reach as much as $25.7 billion by 2030, up from $1.72 billion in 2024.

By early 2026, over 173 AI-originated molecules were in clinical programs, up from just 24 in late 2023.

AI-guided molecular screening is achieving hit rates of 22–46%, compared to the 1–2% of traditional methods.

Discovery-to-clinic timelines that once stretched 4–6 years are being compressed to under 18 months — and these aren’t hypothetical projections. They’re happening in real labs, right now.

AstraZeneca reports AI has enabled more than 50% faster target drug design. Pfizer used an AI-driven mRNA algorithm to increase vaccine production by over 20,000 doses per batch. GSK consolidated 8+ petabytes of trial data into a unified data lake, slashing query times from nearly a year to 30 minutes. Eli Lilly launched “TuneLab” — opening its proprietary AI suite to external biotechs — while committing $1 billion to a collaboration with NVIDIA’s BioNeMo platform. Formation Bio cut clinical trial timelines by ~50%. Cipla reduced manufacturing costs by 26%. Johnson & Johnson cut unplanned factory downtime by 50% with predictive maintenance AI.

And then there’s Insilico Medicine, which may have fired the most significant shot yet. Its AI-designed drug rentosertib (ISM001–055), targeting idiopathic pulmonary fibrosis, reached Phase IIa trials in under 30 months — roughly half the traditional timeline.

The results, published in Nature Medicine, marked the first time an end-to-end AI-discovered drug demonstrated therapeutic benefit in humans. Takeda and Nimbus Therapeutics added another proof point: zasocitinib, an AI-designed molecule for plaque psoriasis, showed efficacy in late-stage trials in late 2025.

The optimist — the Boomer in this story — looks at all of this and sees the dawn of a new era. AI Phase I success rates are running at 80–90%, nearly double the historical industry average of 40–65%. Near-term R&D cost reductions of 15–22% are already being realized.

Insilico’s rentosertib isn’t just one drug; it’s a proof-of-concept for an entirely new R&D engine. The pipeline is filling. The wins are real.

The Algorithm Meets the Human Body

But here’s where the story turns.

Back in Oxford — and in San Francisco, and in London — the AI-first companies that promised to revolutionize drug discovery are running into a wall that no algorithm has yet learned to climb: the human body.

Exscientia’s DSP-1181 was discontinued. Its A2A inhibitor EXS21546 was dropped for failing to demonstrate a viable therapeutic index. The company was acquired by Recursion Pharmaceuticals in 2025 — itself a company whose lead candidate for cerebral cavernous malformation (REC-994) has shown inconclusive efficacy data.

BenevolentAI famously identified baricitinib as a COVID-19 treatment in under three days — a genuine triumph — but its lead eczema candidate failed in Phase II trials in 2023, forcing a painful strategic pivot.

The overall drug development failure rate? Still approximately 95% between Phase I and regulatory approval. AI has not moved that needle.

This is the climax — the moment where the hype meets the hard truth. And understanding why requires looking beyond the technology itself.

The organizational failures run deep. R&D, IT, regulatory, and quality teams operate in silos, preventing coherent AI strategy. Incentive structures reward pushing candidates forward on efficacy signals while downplaying toxicity risks — a dynamic that has sent countless molecules to expensive, late-stage deaths. Toxicity accounts for an estimated 56% of clinical setbacks.

The governance vacuum made things worse. For years, the industry deployed AI models without standardized frameworks for validation or auditability. The FDA only began providing formal guidance — its “Guiding Principles of Good AI Practice in Drug Development” — in January 2026.

Infrastructural fragmentation is equally crippling. Data sits in legacy silos, making the holistic analysis AI promises nearly impossible. GSK’s 8-petabyte unified data lake is the exception, not the rule.

Technologically, the “black box” problem looms large. When an AI model recommends a molecule and can’t explain why, regulators and scientists are left in the dark. Early AI-first companies managed dozens of disconnected, narrow models that produced contradictory predictions — a fragmented “hydra” that proved unsustainable.

Then there are the ethical dimensions that rarely make press releases. Training datasets frequently under-represent diverse patient populations — by race, ethnicity, gender, age — meaning AI-optimized drugs may work better for some groups than others. Patient data used to train these models must navigate HIPAA and GDPR compliance, and the line between innovation and exploitation can blur quickly.

Finally, change management — perhaps the most underestimated challenge. The cultural gap between computational scientists and wet-lab biologists is real and wide. Resistance, distrust, and fear of displacement can quietly kill an AI initiative that looks perfect on paper.

What the Survivors Know

The companies navigating this landscape most successfully share a few hard-won lessons.

They invest in data quality before deploying models — embracing FAIR data principles as a prerequisite, not an afterthought.

They establish cross-functional governance structures bringing R&D, IT, regulatory, and quality teams together from day one.

They validate AI predictions with physical experiments — a “dual-track” approach keeping human scientific judgment in the loop.

They engage regulators early and transparently. And critically, they don’t over-promise.

The companies that have stumbled hardest are those that sold AI as a magic wand to investors, then discovered that queuing up molecules faster just means failing faster.

The Doomer in this story — the pessimist — looks at the 95% failure rate and sees a cautionary tale about hype. AI has optimized the cheapest, easiest part of drug discovery while leaving the hard, expensive, human part untouched.

Phase II success rates for AI-designed drugs are approximately 40% — no better than the historical average. Faster pipelines, the Doomer argues, are just faster pipelines to disappointment.

Both perspectives contain truth. The resolution isn’t a verdict — it’s a posture.

AI in drug discovery is neither the salvation its boosters promised nor the expensive distraction its critics fear. It is a powerful tool being deployed into a genuinely complex system, by organizations still learning how to use it responsibly.

The algorithm that couldn’t save Exscientia’s OCD drug may yet help save someone’s life — if the humans wielding it learn from what went wrong.

Sources

BioSpace — AI Is Changing Pharma’s Bottom Line Now

STAT News — Is this the beginning of the AI-in-drug-discovery era, or the beginning of the end?

Nature — The AI drug revolution needs a revolution

Drug Discovery Trends — Why AI alone won’t resolve drug discovery challenges

The Globe and Mail — Revolution, interrupted: Why AI has failed to live up to the hype

BCG — Reigniting Biopharma’s Research Engine

McKinsey — How Pharma Is Rewriting the AI Playbook

Intuition Labs — Measuring AI ROI in Drug Discovery

Intuition Labs — AI Pharma & Biotech Drug Discovery Case Studies

Drug Target Review — Fixing Failed Drugs: AI Solutions for Toxicity

FDA — Artificial Intelligence for Drug Development

Insilico Medicine — 2025 Annual Announcement

Fortune — Recursion Pharmaceuticals AI Drug Discovery

Fierce Biotech — Exscientia Chops Pipeline

2 Minute Medicine — AI-Designed Drugs Hit 90% Phase I Success Rate

Drug Discovery Trends — Six Signs AI-Driven Drug Discovery Trends

Clinical Leader — AI and Pharmaceutical Development: Ethical Framework

SCW AI — AI in Pharma